Abstract
Predicting the binding of viral peptides to the major histocompatibilitycomplex with machine learning can potentially extend the computationalimmunology toolkit for vaccine development, and serve as a key component in thefight against a pandemic. In this work, we adapt and extend USMPep, a recentlyproposed, conceptually simple prediction algorithm based on recurrent neuralnetworks. Most notably, we combine regressors (binding affinity data) andclassifiers (mass spectrometry data) from qualitatively different data sourcesto obtain a more comprehensive prediction tool. We evaluate the performance ona recently released SARS-CoV-2 dataset with binding stability measurements.USMPep not only sets new benchmarks on selected single alleles, butconsistently turns out to be among the best-performing methods or, for somemetrics, to be even the overall best-performing method for this task.